September 2, 2026 — Anthropic is revising its data retention policy for enterprise customers after businesses raised concerns about how their data would be stored and reviewed when using the company’s most advanced AI models.
The change reverses a policy introduced in June that required prompts and outputs involving certain “covered models” to be retained for 30 days. Anthropic said the retention was intended to support safety monitoring and help detect potential misuse of increasingly capable AI systems.
New safeguards give companies greater control
Instead of relying on the blanket 30-day retention requirement for enterprise users, Anthropic plans to introduce a new system called Enterprise Frontier Safeguards.
The framework is designed to give businesses greater control over how their data is stored and monitored. Anthropic says customers will be able to use automated safety monitoring without requiring Anthropic employees to review their data. The safeguards are expected to be rolled out in phases, with broader availability planned for later this year.
The company developed the revised approach after receiving extensive feedback from enterprise customers. Anthropic has been working with businesses that handle sensitive information to address concerns around privacy while maintaining mechanisms to detect potentially dangerous use of its frontier models.
Balancing privacy with AI safety
Anthropic's original policy reflected a growing challenge for AI companies: more powerful models can create new security risks, but monitoring their use can require access to user data.
The company has argued that retaining limited data can help identify misuse and improve safeguards. Its safety roadmap has also acknowledged the tension between zero-data-retention arrangements and the need to detect abuse as AI capabilities become more advanced.
The revised enterprise approach attempts to address both concerns by giving organizations more control over their information while maintaining automated safety protections.
Policy remains for some advanced models
Anthropic's broader data-retention rules are not disappearing entirely. Its current policy says prompts and outputs involving designated covered models are retained for 30 days as part of its safety work. The policy applies across platforms where those models are offered, while other models remain under their existing terms.
For enterprise customers, however, the planned Frontier Safeguards represent a significant shift toward customer-controlled data handling.
The move comes as companies increasingly scrutinize how AI providers handle confidential business information. For Anthropic, the challenge will be to demonstrate that stronger privacy protections can coexist with the safety monitoring it considers necessary for increasingly capable AI systems.
